Distinguishing schizophrenia and bipolar disorder through disorder-specific network revealed by contrastive machine learning

Schizophrenia (SCZ) and bipolar disorder (BD) are severe psychotic disorders with overlapping clinical manifestations, leading to high rates of misdiagnosis. This study aims to identify disorder-specific neurophysiological biomarkers using electroencephalography and contrastive machine learning to improve differential diagnosis. Resting-state electroencephalography was recorded from 52 patients with BD, 65 with SCZ, and 75 healthy controls. Temporal variability networks were constructed using sample entropy. Contrastive variational autoencoders decomposed these networks into components shared with healthy controls and components specific to each disorder. Based on disorder-specific components, predictive models for clinical symptoms were constructed. Additionally, spatial pattern network filters were implemented to extract discriminative features for the classification of BD and SCZ patients. Here we show pronounced differences in disorder-specific network components between SCZ and BD, especially in frontal-central/parietal connectivity, which were not discernible in the original or shared networks. These disorder-specific components correlate significantly with clinical assessment scores and support predictive modeling of symptom severity. By applying spatial pattern network filters to the disorder-specific components, we achieve 96.154% accuracy in distinguishing SCZ from BD, substantially surpassing conventional approaches. This integrative framework, combining dynamic network analysis with contrastive machine learning, provides a powerful methodology for extracting neurophysiological biomarkers and paves the way for biologically grounded diagnostics in psychotic disorders. Schizophrenia and bipolar disorder are serious mental illnesses that can look very similar, making them hard for doctors to tell apart. This study used an objective brain index (EEG) combined with an artificial intelligence method to find differences between the two conditions. The researchers analyzed brain activity patterns and isolated features unique to each disorder. They discovered that these unique patterns were linked to patients’ symptoms and could predict how severe those symptoms were. Most importantly, the method distinguished between the two disorders with over 96% accuracy. This work could lead to a more reliable, biology-based tool to help doctors diagnose patients correctly, reduce misdiagnosis, and guide more personalized treatment decisions in the future. Jiang, Ye et al. combine resting-state EEG and contrastive variational autoencoders to identify disorder-specific neurophysiological biomarkers in schizophrenia and bipolar disorder. The approach achieves over 96% accuracy in distinguishing the two disorders by revealing distinct frontal-central/parietal connectivity patterns.